Krea2 Turbo 2-Step Distill LoRA Final: 4x Faster, Detail Within 92% of the 8-Step Teacher
TimeTruth2490 · reddit · 2026-10-11
The final checkpoint (chk51195) of the community-trained Krea 2 Turbo 2-step distillation LoRA cuts denoising from 8 to 2 steps—56.9s to 14.3s at 1024x1024 (4x faster).
Key numbers and findings:
- Detail retention: fine-texture energy reaches 0.92-1.02x of the teacher across 12 trained resolutions (stock Turbo at 2 steps: 0.40-0.65x); a blind rubric finds only 1 of 352 prompt-conditioned items missing.
- Saved steps become resolution: renders up to 2048x2048, Krea's published recommended max; beyond that the stock model duplicates subjects regardless.
- Limits: small in-frame subjects (faces in crowds) can ghost or smear—use the companion 4-step LoRA when quality matters.
- Training: rank 64 on 228 modules, 23,561 teacher trajectories, 51,195 samples, 32 days on a single RTX 3090, 40+ recipe adjustments; final weights from a 600-sample LR-to-zero anneal, not a running average.
- Drop-in: plain LoRA, stock Euler, works with diffusers/ComfyUI/MLX, no custom nodes.
The project is closed; the author has moved on to a 3-step adapter for Qwen-Image-2.1-Turbo.
More from Multimodal
- Higgsfield casts viral AI influencers in action film made with Katana video model — CurieuxExplorer · 2026-10-11
- Creator shows off a fully AI-generated music video made from one prompt — flngr · 2026-10-11
- Kling-generated soul music video shows off motion-controlled singing and dance — Fearless_Ad_7811 · 2026-10-11
- ZPix Adds Qwen-Image 2.1 Turbo with Checkerboard-Fix VAE for Local Image Generation — SamuelTallet · 2026-10-11
- CUT debuts: build scenes with AI characters, voices, and a timeline — rileygstaliger · 2026-10-11
- NVIDIA's GATOR turns casual photos into simulation-ready 3D objects with agentic refinement — AjayMandlekar · 2026-10-11